This change allows .experts.gate_proj / .up_proj / .down_proj tensor names to each
be used for both quantized (i.e. nvfp4 and mxfp8) and non-quantized (bf16) models.
Previous to this only non-quantized models used that tensor naming scheme.
In the PS output, expose the scheduler selected size (clamped by model context size) instead of always reporting the model max context. This will help provide a hint to clients to keep the context size below this value to avoid paging and poor performance on smaller VRAM systems.
Bump MLX to the latest selected upstream ref and update the MLX/imagegen
wrappers and tests for the new API behavior.
Fix the CUDA MLX archive so runtime NVRTC kernels work after deployment:
package CUTE/CUTLASS headers, include the CUDA runtime header closure, and
stage a coherent CUDA-toolkit-matched CCCL tree instead of MLX's fetched CCCL
for CUDA payloads. The previous archive could build successfully but crash at
runtime due to missing or incompatible JIT headers.
The heuristic schedule grew the draft toward a fixed cap on acceptance alone,
maximizing accepted-tokens-per-step rather than throughput, and on a
steep-forward target it regressed below no speculation. Replace it with an
engine-level controller that drafts the depth maximizing
committed-tokens-per-wallclock from live per-position acceptance and persisted
per-width forward cost, with no draft-length cap; the heuristic schedule and
the OLLAMA_MLX_MTP_* env vars go with it.
Acceptance took two blocking evals per round: one to read the accepted mask,
then a second for the bonus or residual token whose graph needed the
host-known rejection point. Sample the residual at every rejection point in
one batched draw alongside the bonus row, so a single eval covers acceptance
and the next token.
Each speculative round ran the target stack twice — once for the current
token's hidden and base logits, once to validate the drafts — capping
throughput below plain decode. Fuse them into one forward over [current,
draft_0..draft_{N-1}], whose hidden rows already line up with the acceptance
math, so the separate base-logits unembed disappears from the drafted path.
Sampler.Distribution built row i as if draftTokens[:i] were appended, leaving
a single-row proposal call with no draft history, so a drafter skipped the
repeat/presence penalties the target's validation applies and re-proposed
penalized tokens. Align rows with the end of the draft chain instead: the
final row sees every draft token, each earlier row one fewer.
Generalize the draft path so a head that maintains a KV cache (EAGLE-style)
and Gemma's read-only single-position assistant both fit one drafter
interface with no per-model branches, and make the committed stream the
drafter's maintenance mechanism — every committed run is reported, the
drafter pairs each draft slot with its look-ahead token and flushes completed
pairs to the draft caches. The draft KV thus stays prefix-cached alongside
the target in every session, drafting or not.
The pipeline and the MTP decoder each owned a decode loop with duplicated
prefill, budget, and emission handling. Split the pipeline into prefill and
decode phases behind a decoder interface, with the decode loop the sole
emitter enforcing the NumPredict budget, and split speculation into a generic
engine that returns the accepted run and a drafter interface that owns only
how proposals are made.
Greedy is a special case of sampled decoding — at temperature 0 the sampler
yields a point mass, so rejection-sampling acceptance reduces to argmax-match
— so collapse the separate greedy, sampled, and serial paths into one. MTP
now honors any temperature, penalty, and top-k/p/min-p setting; logprobs
remain the only gated feature.
Adding/Multiplying a tensor by a scalar w/ a different data type
can cause the tensor to be promoted and cause performance issues.
This change adds several guards against over-promotion.
The batched MTP accept paths advance the cache by the whole accepted run
before streaming it to the client. If the stream was cancelled partway
(e.g. the caller disconnects), the loop returned before recording the
remaining accepted tokens, leaving the cache offset ahead of
session.outputs. close() then indexed the token log past its end and
panicked with a slice-bounds error.
Record the whole run to session.outputs before streaming any of it, so a
cancelled stream can no longer desync the cache from the token log.
The same bug is present on main, with identical mechanics: the accept
paths there commit the cache to before+accepted and then stream in a loop
that returns on cancellation before recording the rest.
Prefill no longer splits its batch at each requested snapshot offset. The
session schedules the pending offsets on every cache before prefill, runs the
forward in full-size chunks, and attaches the captured snapshots to the trie
afterward. Offsets the prefill never crosses (it leaves one token for decode
seeding) are dropped instead of materializing a node for tokens never written,
and snapshots from an abandoned prefill are released on session close.
Speculation used a parallel hierarchy of wrapper cache types that shadowed
the live caches and reconciled against them on commit. Replace it with
snapshot/restore on the live caches themselves: a cache snapshots itself as
a write crosses each offset, and the runner commits a batched draft by
restoring to the accepted count. The wrappers and the comparison plumbing
around them are gone.
Snapshots are lazy. A KV or rotating capture indexes into the live buffer and
owns no memory until a destructive write forces a copy-out, so rejecting a
draft is free.
Recurrent layers now validate in the same batched pass rather than falling
back to serial. A gated-delta layer reports its interior split offsets and
hands back the recurrent state at each one, which the cache records as a
snapshot.
CausalConv1D and GatedDelta now run their scan in segments cut at optional
WithSnapshotSplits offsets and return the recurrent state at each boundary
instead of just the final state. The output is identical to the unsegmented
scan; segmenting only adds a few kernel launches, not extra recurrence compute.
This lets a batched forward capture interior recurrent state without re-running
the scan, which the cache will use for speculative validation rollback points.
RecurrentCache.Put and the Qwen3.5 layer now thread the boundary-state slices,
committing the final entry as the live state.
cache.go had grown to hold every cache kind. Move KVCache (and its
speculative wrappers) to kvcache.go and RotatingKVCache (and its
sliding-window mask applier) to rotating.go, leaving cache.go with the
shared interfaces and the Speculation transaction. Pure relocation;
no behavior change.
Work that panics on the locked MLX worker goroutine was recovered and
re-raised on the caller, so the printed trace pointed at the re-panic
site in this package rather than the code that actually panicked.
Capture the worker stack at recovery and carry it through a value that
implements error, so the runtime prints the original location in the
fatal trace.
* llama-server followups
Misc fixes for #16031
- Add back dropped ROCm build flag for multi-GPU support on windows
- Fix amdhip64_*.dll version detection for "latest" selection
- Fix embeddings API for consistent normalize behavior with prior versions
* ci: set up for automated llama.cpp update testing
* reduce batch for fa-disabled, and constrained vram
* mlx: fix v3 load bug on m5
Imagegen was incorrectly loading v3 first. This DRYs out the loading code so imagegen gets the same new v4/v3 selection logic.
* fix reload bug on embedding models
* bump version
* steer user how to enable iGPU when disabled
* broad lint fixes to sidestep CI scope glitch
* runner: Remove CGO engines, use llama-server exclusively for GGML models
Remove the vendored GGML and llama.cpp backend, CGO runner, Go model
implementations, and sample. llama-server (built from upstream llama.cpp via
FetchContent) is now the sole inference engine for GGUF-based models.
(Safetensor based models continue to run on the new MLX engine.) This allows
us to more rapidly pick up new capabilities and fixes from llama.cpp as they
come out.
On windows this now requires recent AMD driver versions to support ROCm v7 as
llama.cpp currently does not support building against v6.
* llama/compat: load Ollama-format GGUFs in llama-server
Squashed from upstream/jmorganca/llama-compat on 2026-04-29.
Source tip: 0c33775d378511a9b3c7f2e3b80eda355511d9f3.
Original source commits:
- 25223160d llama/compat: add in-memory shim so llama-server can load Ollama-format GGUFs
- 7449b539a llm,server: route Ollama-format gemma3 blobs through llama/compat
- 436f2e2b1 llama/compat: make patch-apply idempotent
- 8c2c9d4c8 llama/compat: extend gemma3 handler to cover 1B and 270M blobs
- 021389f7b llama/compat: shrink clip.cpp injection from 18 lines to 1
- 61b367ec2 llama/compat: shrink patch to pure call-site hooks (34 -> 20 lines)
- 36049361c llama/compat: simplify shim (gemma3-tested)
- 8fa664865 llama/compat: add qwen35moe text handler
- db0c74530 llama/compat: add qwen35moe vision (clip) support
- 2a388da77 llama/compat: split shared infra into a util TU
- 9a69a17dc llama/compat: document non-public API dependencies
- d0f38a915 llama/compat: add gpt-oss and lfm2 handlers
- 086071822 llama/compat: add mistral3 text handler (vision TODO)
- 63bde9ff7 llama/compat: add mistral3 vision (clip) support
- 3a57b89d5 llama/compat: apply LLaMA RoPE permute to mistral3 vision Q/K
- 99cb87439 llama/compat: add qwen35, gemma4, deepseek-ocr handlers
- 2c7850dba llama/compat: add nemotron_h_moe handler (latent FFN + MTP skip)
- 9e3b54225 llama/compat: add llama4 text + clip handlers
- 034fee349 llama/compat: add gemma4 clip handler (gemma4v projector)
- 9945c5a93 server: remove dhiltgen/* compat redirect table
- 5d4539101 llama/compat: rewrite gemma4 tokenizer model to BPE
- 7e0765327 llama/compat: add glm-ocr text handler + text-loader load-op hook
- f1bd1a25a llama/compat: add glm-ocr clip handler (glm4v projector)
- 4b5cf3420 llama/compat: collapse text-loader hook back to one new patch line
- eb4ecf4fc llama/compat: extend gemma4 clip handler to gemma4a (audio)
- a23a5e76f llama/compat: fix gemma4a per-block norm tensor mapping
- cd2dcaff4 llama/compat: add embeddinggemma handler
- 1ce8a6b26 llama/compat: add qwen3-vl + qwen2.5-vl handlers
- fd98ffa1e llama/compat: add gemma3n + glm4moelite handlers
- cc7bdf0bc llama/compat: handle null buft in maybe_load_tensor
- 0c33775d3 llama/compat: disable mmap when load_op transforms text-side tensors
* refine implementation
* ci: fix windows MLX build
* ci: fix windows llama-server build
* ci: fix windows rocm build
* ci: windows mlx tuning
Shorten long-tail on build, and get OllamaSetup.exe back under 2g limit
* ci: fix windows dependencies
* win: fix dependency gathering
* disable openmp
* win: arm64 cross-compile build
also DRY out CI steps
* scheduler improvements
* ci: improvements from #15982
* win: favor ninja for faster developer builds
* win: fix build
* win: fix arm64 cross-compile
* win: avoid spaces in compiler path
* misc discovery fixes, and bos handling
* lint fixes
* win: fix arm cross-compile build/CI bugs
* llama.cpp update
* win: handle multiple CRT dirs
* vulkan: add windows iGPU detection
* fix creation bugs for patched models, other refactoring work
* tune batch size for better performance
* ci and lint fixes
* fix repeat_last_n bug
* build: revamp build for better developer UX
* amd, sampler, qwen3next fixes
* version bump
* fix mlx build
* revamp GPU discovery
Scanning the output of llama-server is turning out to be too error prone across
llama.cpp updates, so this switches to a thin dynamic library load against the
bundled GGML libraries so more details can be gathered from the API.
* version bump
* missing file
* ci: fix cache miss on rocm build
* refine vulkan dep handling
* fix ps reporting bug on full GPU load
* improve cmake wiring for customized local builds
* version bump
* docker build arg cleanup
* improve windows exit error logs
* fix community gemma4 support and ci flakes
* fix mlx unit test
* tighten up ps logic to avoid double counting fit log lines
* version bump
* fix ps view for full gpu layer offload
* add MTP wiring for llama-server and create with GGUFs
* pick best template by capabilities
* version bump
* ci: harden apt repos
* remove unused cpu core discovery
* adjust batch default logic to reduce OOMs
* support larger tool calls
* fix audio support, template show
* qwen35 mtp patch support
* flesh out dtypes
* rocm deps
* version bump
* lint fix
* block broken gfx1150 on windows
* fix qwen3.5 moe mtp tensors in patch
* mmproj oom fallback and vulkan on by default
* qwen MTP compat fix
* version bump
* ci: fix WoA cross-compile
* ci: workaround ui tool in cross-compile
* version bump
* win: enable OpenMP for CPU builds
* build: improve developer UX
* ci: windows path workaround for CPU build
* win: fix WoA dependencies
* win: fix large offset reads for mmproj patched loads
* version bump
* fix vulkan dup detection
* add OLLAMA_IGPU_ENABLE and largely disable iGPUs by default
* opt-in MTP, win large offset, integraton fixes
* fix unit test scheduler interaction hang
* fix multi-gpu filtering
* version bump
* review comments
* fix thinking level
* fix linux rocm ordering and granite 3.3 template
* version bump
* ci fix - non-shallow MLX checkout
* bypass linux sysfs unit test on windows
---------
Co-authored-by: jmorganca <jmorganca@gmail.com>
This change updates the show API for MLX models to:
* display the correct quantization in mixed precision models
* not display the global_scale scalar value
* not duplicate the `tools` capability
Split the gated-delta Metal/CUDA kernels' dtype template into separate
input (InT) and state (StT) types so activations can stay in bf16/fp16
while the accumulated delta state stays in float32. Allocate the delta
state and qwen3_5's no-cache zero state in float32 to match.
Previously the draft architecture was hardcoded to
Gemma4AssistantForCausalLM. Read it from the draft model's config so
any draft architecture can be packaged.
This reverts commit 98e26b8c37.
The DFlash integration is too invasive to keep at this stage: it
threads DFlash-specific logic through the pipeline, base model
interfaces, and the cache layer. The recurrent cache also now
has qwen3.5 model-specific code. Revert it now and reintroduce
the self-contained, generally-useful pieces (YaRN RoPE DRY-out, draft
architecture autodetection, gated-delta fp32 state) as separate
follow-up commits.
* ci: speed up release builds
This should help speed things up for release. It also will help
speed up local developer builds a little.
* ci: dedup linux build steps and optimize
* review comments
This change adds dflash block diffusion speculative decoding to the MLX runner. Included in this change:
support for qwen3.6 moe/dense speculative decoding
draft model recurrent cache playback
RoPE/YaRN changes (DRY out the laguna/dflash MoE YaRN implementation)
support for greedy sampling / leviathan/chen sampling
* mlx: rework the MLX sampler
Replace the MLX sampler transform chain with an explicit distribution pipeline that applies:
1. penalties
2. top-k
3. temperature/softmax
4. top-p
5. min-p
6. normalize
7. categorical
The common top_k path now keeps sparse [B,K] token ids/probabilities on GPU instead of carrying full-vocab
scores, and sampled MTP reuses those draft/target distributions for acceptance, bonus, and residual sampling.
This change also fixes the seed parameter so that temperature sampling and sampled MTP are reproducible.
The MLX runner now routes model work through a locked worker thread. Status also used that worker only to sample memory, so a scheduler health ping could sit behind long prefill or generation until its 10s context expired, causing /v1/status to return 500 and the server to treat the runner as unhealthy.
While Metal doesn't change VRAM reporting, CUDA does. Cache the last memory sample and make status perform only a short best-effort refresh. If the worker is busy, status returns the cached value while a single background refresh continues and updates the cache when the worker becomes available. The in-flight guard and lifecycle context keep this from spawning unbounded refreshes while preserving live VRAM refresh behavior for CUDA.
Fixes#16081
* mlx: refined model push behavior
Refine the algorithm for parallel push of safetensors based models to get
better reliability and throughput.
* review comments, hardening, and performance tuning for slow links
* review comments
This change adds support for MTP (multi-token prediction) speculative decoding for the
gemma4 model family.
It includes:
* support for importing safetensors based gemma4 draft models with `ollama create`
* a new DRAFT command in the Modelfile for specifying draft models
* a --quantize-draft flag for the ollama create command to quantize the draft model
* cache support for speculation
* changes to the rotating cache to be able to handle MTP correctly
* sampling support for draft model token prediction
---------
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
* Update MLX and MLX-C
* Run MLX CGO work on a locked OS thread
MLX now relies on OS-thread-local execution state for streams, encoders, and caches. Add an mlxthread executor backed by runtime.LockOSThread and route runner initialization, model load, inference, status memory reads, and cleanup through the worker so Go goroutine migration cannot split MLX state across native threads.
Also stop caching default MLX streams before the runner owns the thread and add worker/threaded MLX regression tests.
* mlx: use common status writer
* mlx: bundle missing libjaccl on arm64
Inspired by #15793
* review comments
* metal: harden for ggml initialization failures
ggml_metal_device_init performs a probe to verify the tensor API compiles. On
some systems this passes, even though kernel coverage isn't complete, which
results in a later crash when compiling the real kernels. This change adds a
single retry if any of the error strings match this failure mode to disable the
tensor API. It also hardens an error case in the Go initDevices to detect
device initialization failures and panic instead of crashing later on a nil
array entry.
Fixes#15734
* review comments
* review comments
* mlx: add laguna model support
* convert: support fp8 safetensors import
Decode HF F8_E4M3 safetensors with block scale companions into GGUF-supported tensor types, and record which output tensors came from FP8 source weights.
Use that source-precision metadata during create quantization: default FP8-sourced GGUFs to Q8_0, keep non-FP8 tensors at their original precision for Q8_0, and promote non-FP8 quantizable tensors to Q8_0 for Q4_K requests.
* ggml: add laguna model support
* server: preserve generate logprobs with builtin parsers
Generate requests were dropping logprob-only chunks whenever a builtin parser buffered visible content. Chat already handled this case, but generate only forwarded chunks with visible response, thinking, or tool-call output.
Keep generate chunks that carry logprobs even when the builtin parser has not flushed visible content yet, and add a regression test that exercises the behavior with a generic thinking parser.
* review comments - perf improvements
* ggml: implement nemotron 3 nano omni
* add poolside integration
* update poolside doc
* adapt to new cache setup
* fix test
* fix test
---------
Co-authored-by: Eva Ho <hoyyeva@gmail.com>
Models build their own attention masks and read K/V directly from
the cache's buffers, which ties them to the cache's storage layout.
That blocks multi-sequence batching — right-padded rows need a
query-padding mask composed onto every model — and rules out
variants like paged attention where K/V isn't one contiguous tensor.
Caches now hand back a per-layer KVHistory holding post-update K, V,
and a MaskApplier that merges the cache's storage restrictions into
the model's logical mask. Models describe their mask in logical
terms; SDPA composes model, padding, and applier contributions and
dispatches to the kernel's causal or no-mask fast path when it can.
KVHistory still exposes K, V, and the composed mask for manual
attention paths (e.g. CUDA prefill at head_dim > 128).
Performance for single-sequence inference is unchanged.
Switch RoPE from the scalar-offset kernel (mlx_fast_rope) to the
array-offset one (mlx_fast_rope_dynamic) so each batch row can start
at its own position. The pipeline tracks the current position locally
and passes it to the model through Batch.SeqOffsets; each model
materializes that slice into an int32 array for the RoPE call.
Single-sequence behavior is unchanged; this is the wiring needed
before the runner can batch independent sequences.
Gives a single extension point for per-call context (positions,
sequence IDs, masks) as multi-sequence batching grows, without having
to churn every model's Forward signature again.
* mlx: Support NVIDIA TensorRT Model Optimizer import
* x/create: support FP8 safetensors import
Decode HF F8_E4M3 safetensors with block scale companions into MLX-importable tensor blobs, including compressed-tensors weight_scale metadata, packed NVFP4 layouts, and mixed-precision tensor headers.
Use that source-precision metadata during create quantization: default FP8-sourced imports to mxfp8, allow source FP8 to target MLX low-bit formats, preserve source-quantized NVFP4 layouts, selectively keep or promote tensors based on their source precision, and detect quantized dtype from mixed-precision safetensors manifests.
* review comments
Register sequences with Add/Remove; each Sample call takes any subset of
registered slots and samples one token per row, appending to each slot's
ring-buffer history. When all slots share Options and penalty rings are
full, one fused transform pass runs over the whole batch via a persistent
pooled history tensor; otherwise calls fall back to per-slot serial
processing indexed against the same pool.
Performance is unchanged for a single sequence, which is all that is
exposed for now.
AppendToken used to concatenate the new token onto the history tensor
and slice it back to RepeatLastN every decode step, churning the graph
shape and reallocating a fresh tensor each call. The stateful penalties
don't care about order within the window, so a fixed-capacity ring with
one SliceUpdate per append keeps the tensor shape constant across
steps.
Move tokenization out of the single GPU processing goroutine and
into each request's HTTP handler goroutine. This allows the next
request's prompt to be tokenized on the CPU while the current
request is executing on the GPU.
